Approaches to modernize the combination drug development paradigm

Types of MTA combinations

Approved MTA combinations

In addition to the MTA–MTA and MTA–endocrine-therapy combinations, ipilimumab and nivolumab are two immuno-oncology agents also approved as a doublet regimen. Rather than targeting aberrant genomic pathways, these monoclonal antibodies (mAbs) inhibit immune regulatory checkpoints, cytotoxic T-lymphocyte-associated antigen 4 (CTLA-4) and programmed cell death protein-1 (PD-1), respectively, producing durable tumor regression in multiple tumor types [5–11]. Mechanistically, combined CTLA-4 and PD-1 blockade demonstrated enhanced treatment efficacy by targeting non-redundant immune pathways [31, 32].

The scientific basis of these nine FDA-approved combinations is founded on proof of resistance mechanisms to an established therapy and/or evidence of synergistic or additive activity in animal models [28, 33–42]. The targeting of the mitogen-activated protein kinase (MAPK) pathway at two key levels is an example of using dual targeted therapy to effectively counteract genetic escape mechanisms. In the treatment of advanced malignant melanoma, combined inhibition of BRAF and its downstream effector MAPK kinase (MEK) led to improved survival outcomes compared with BRAF inhibition alone. The doublet regimen prevents MAPK pathway activation, which is the most common mechanism of acquired resistance to BRAF inhibitors [27, 29, 38, 39, 43, 44]. Notably, in these nine approved combinations, MTAs are used at or near their single-agent recommended dose, without substantial increase in toxicity. Additionally, in seven out of the nine combinations—with the exceptions of lenvatinib and everolimus, and nivolumab and ipilimumab—established predictive biomarkers are utilized for molecularly based patient selection [22–30].

Lessons learnt from unsuccessful MTA combinations

Failure to show benefit at the phase III stage is particularly disappointing during drug development given the immense financial cost and human resources involved. Importantly, a large number of patients may have been exposed to harm or unnecessary treatment. An example is the combination of anti-epidermal growth factor receptor (EGFR) antibodies (cetuximab or panitumumab), anti-vascular endothelial growth factor (VEGF) antibody (bevacizumab), and chemotherapy in metastatic colorectal cancer. Both anti-EGFR and anti-VEGF mAbs have established antitumor activity in combination with chemotherapy in this setting [46–48]. A large body of nonclinical data demonstrated synergism produced by combined EGFR and VEGF blockade and a small phase II study in irinotecan-refractory patients demonstrated clinical benefit [49–51]. However, in two phase III trials, which together included over 1700 patients, the addition of cetuximab or panitumumab to bevacizumab and chemotherapy in the frontline setting unexpectedly resulted in significantly shorter progression-free survival, including in the KRAS wild-type subpopulation, and toxicity was also increased [52, 53]. The reasons behind this discouraging outcome are not known. Discontinuation rates secondary to toxicity were similar in both arms in one study [53]. The investigators postulated that unfavorable pharmacokinetic and pharmacodynamic interactions between the anti-EGFR and anti-VEGF antibodies may have occurred, leading to the blunting of the therapeutic effect of each agent [52, 53]. This example serves as a reminder that although combining drugs with proven mechanisms of action is an attractive and logical strategy, carefully designed early clinical trials with comprehensive pharmacokinetic and pharmacodynamic parameters are essential to understand drug interactions and to provide proof of concept.

Selecting and prioritizing combinations: a systematic approach to drug discovery

Currently, there are approximately 100 approved anticancer drugs and, according to the Pharmaceutical Research and Manufacturers of America’s 2015 report, approximately 1200 new anticancer drugs are in development in the United States, amounting to approximately 845,000 possible pairings and an exponential number of higher-order combinations [54]. This, in addition to the expanding number of potential drug targets, necessitates systematic and efficient methods of drug discovery.

Increasingly, industry and academia are utilizing various methods of high-throughput screening, which leverage laboratory automation to simultaneously assay the biological activities of a vast number of compounds [55–59]. Indeed, unbiased chemical screening may uncover unexpected interactions, likely attributable to previously unknown interconnected cellular signaling pathways [60, 61]. For example, in an attempt to identify therapeutic partners for the Bruton’s tyrosine kinase (BTK) inhibitor ibrutinib, a high-throughput screen study in diffuse large B-cell-lymphoma cell-line models demonstrated impressive combinatorial activity with a range of mechanistically distinct drug classes, which may warrant further investigation [55]. Additional technological advances include in silico modeling methods to facilitate large-scale genome-wide identification of candidate synthetically lethal genes as new drug targets and to predict drug response [62]. Computational network-based algorithms can also systematically analyze gene regulatory and signaling pathways to mechanistically define genetic determinants of disease and establish new therapeutic targets [63]. Furthermore, ex vivo testing in cell culture models derived from patient samples with acquired resistance may prove to yield more robust and predictive tumor models than existing models for therapeutic testing [64].

Complementary to these novel approaches, data-sharing efforts are imperative to promote scientific collaboration. Examples include publically available data repositories that catalogue protein–protein interactions and biological pathways such as Pathway Commons and Database of Interacting Proteins [65, 66]. The US National Cancer Institute (NCI) recently launched Genomic Data Commons, an interactive data-sharing platform enabling the import and harmonization of genomic data from multiple research programs using standardized bioinformatics pipelines [67, 68]. Additionally, large annotated cell-line drug-screen libraries, such as the Cancer Cell Line Encyclopedia, Genomics of Drug Sensitivity in Cancer Project, and NCI60 (the US National Cancer Institute 60 human tumor cell line anticancer drug screen), are publically accessible datasets correlating drug sensitivity with detailed genomic data and thus serve as rich resources for researchers [69–74].

Quality of nonclinical data

Nonclinical studies of drug combinations often report synergy without appropriate evaluation of the combination index (Box 1) [75]. Additionally, low rates (11–25 %) of reproducibility of published laboratory data, including those from high-impact journals, have been identified, despite attempts to recreate the experimental environment, suggesting limitations in the validity of scientific findings and probable publication bias [76, 77]. Improving the reliability and predictive value of nonclinical studies is fundamental to successful clinical translation. Box 2 outlines important issues to consider when designing the aims, experimental conditions, and parameters of nonclinical studies. At a minimum, benchmarks for nonclinical studies prior to consideration of clinical testing should include validation of a robust scientific hypothesis via demonstration of mechanism of action, observation of objective synergistic or additive antitumor activity, and acceptable safety at clinically achievable drug concentrations.

Novel dose-finding strategies

Traditional rule-based dose-escalation trial designs, which rely on the premise of dose-dependent toxicity, may have considerable limitations in defining the biological optimal dose and schedule of combination MTAs. Unlike cytotoxic chemotherapy, the relationship between dose, toxicity, target inhibition, and efficacy is less predictable with MTA therapy [78]. Pharmacological interactions of drug combinations can also impact on dose effect. Furthermore, the assessment of dose-limiting toxicity (DLT), a key determinant of the maximum tolerated dose (MTD) and the recommended phase II dose, can be problematic for phase I MTA trials. The traditional DLT window of observation—typically the first cycle of treatment—used in trials of cytotoxic agents may not be adequate, as MTAs may display delayed toxicity due to their relatively long half-life and chronic dosing schedule. There is also a lack of consensus on the definitions of DLT, and these have been found to be widely heterogenous across MTA clinical trials [79].

Adaptive Bayesian model-based designs may be better placed to cater for the complex variables associated with combination MTAs, by incorporating pre-study probability of toxicity and updating such probability with real-time adverse event data to inform dose-escalation decisions [80–83]. In simulation studies, adaptive designs were found to maximize the number of patients treated at or near the MTD compared with rule-based designs [84, 85]. Importantly, adaptive designs allow prospective modifications on aspects of the trial as the data evolve, offering greater flexibility to researchers. Additionally, multiple methods have been proposed using both toxicity and efficacy as endpoints [86–88]. For example, the zone method describes a parallel phase I/II design that uses initial rule-based dose escalation and subsequent Bayesian adaptive randomization to enable simultaneous evaluation of the safety and efficacy of multiple dose combinations, such that sample size can be reduced and more patients are treated with the higher-efficacy dose levels [86]. However, one challenge of adaptive designs is the requirement for continual biostatistical modeling, which may have affected their uptake in the past, although the application of these designs is increasing with time [84, 89, 90]. An operationally hybrid approach is the modified toxicity probability interval design, which obviates the requirement for computational modeling by utilizing an up-and-down dose-assignment scheme conceptually similar to a rule-based algorithm but guided by Bayesian models [91, 92]. This design has been used in a number of phase I clinical trials [93, 94].

Currently, there is no preferred dose-escalation design for combination MTAs. The choice of the most appropriate dose schedule selection and dose-finding method should be informed by knowledge of the nonclinical and clinical pharmacology of the agents of interest and based on consultation between experienced clinical researchers, sponsors, and statisticians. Comprehensive pharmacokinetic evaluation and pharmacodynamic assessment of tumors in early phase trials are vital to assess for target modulation and to mechanistically characterize on- and off-target toxicities. These are particularly pertinent in combination studies to delineate individual drug effects and to evaluate pharmacological interactions. Multiple targets may also need to be cross-examined to assess network inter-dependencies in identifying resistance mechanisms.

Integrating the immune landscape

Combinations of novel immune checkpoint and co-stimulatory molecules are actively being evaluated for additive or synergistic effects. Additionally, as multiple oncogenic pathways can foster immunosuppressive microenvironments, immuno-oncology agents are also being investigated in combination with MTAs, with pending efficacy data [12–17]. Although validated predictive biomarkers are presently lacking, emerging immune monitoring techniques are giving insights into the interactions between tumor antigen profiles, the microenvironment, and the immune response, guiding potentially personalized immuno-oncology strategies [95]. However, the unique properties of immunotherapies present multiple challenges to clinical trial design, demanding careful forethought on dosing and schedule decisions, patient selection, toxicity assessment, pharmacokinetic/pharmacodynamic monitoring, and choice of endpoints. For example, immuno-oncology agents often do not reach a MTD in dose-escalation studies and exhibit distinctive patterns of tumor response and immune-related toxicities, which are not necessarily dose-dependent. Furthermore, nonclinical studies of immuno-oncology agents are technically difficult owing to species-specific differences in immune response and the lack of reliable models [96].

Improving the efficiency of drug development: a shift from the traditional three-phase approach

The desire to expedite drug development has challenged the conventional three-phase clinical trial paradigm, in which safety and efficacy objectives are mandated at each distinct stage to fulfill regulatory requirements. Indeed, the time period from first-in-human testing to regulatory approval for oncology drugs typically averaged 7–9 years, substantially longer than agents of other therapeutic classes [45, 97]. Novel designs including seamless phase I/II, II/III clinical trials and large dose-expansion cohorts in phase I trials are increasingly utilized to streamline clinical development, obviating lengthy pauses between trials [98–100]. For example, the anti-PD-1 mAb pembrolizumab received accelerated approval by the FDA in September 2014 for use in metastatic melanoma, based on efficacy data from 173 patients enrolled onto the expansion cohort of a phase Ib trial (KEYNOTE-001), a mere 3 years after clinical development began [8]. Dose-expansion cohorts are often conducted in phase I trials for anticancer agents given alone or in combination that demonstrate compelling early signals of activity. These “tails” in phase I trials, typically in the form of disease-specific or biomarker-specific cohorts, apply intermediate endpoints, such as objective response rate, to support accelerated approval [99]. However, the use of these strategies requires clearly defined goals, pragmatic protocols, and statistical design with the flexibility to respond to evolving data and objectives, independent oversight, and a commitment from researchers, commercial sponsors, and regulators to work in tandem [99]. Getting these factors right is important to strike a balance between protection of patient safety and interests and therapeutic innovation.

Targeting tumor heterogeneity: genomically informed clinical trial designs and biomarker development

Multiple umbrella (histology-specific) and basket (histology-agnostic, aberration-specific) clinical trials are currently ongoing, incorporating genomic testing results for treatment assignment to genotype-matched therapies and combinations (Lung-MAP, NCT02154490; BATTLE, NCT00409968; NCI-MATCH, NCT02465060; and My Pathway, NCT02091141) (Box 1) [101]. The size and scope of these umbrella and basket trials can vary based on complexities and heterogeneities of tumor genotypes under evaluation. In some cases, adaptive Bayesian model-based designs can be embedded in these clinical trials to dynamically test multiple hypotheses, doses and schedules (Fig. 1). The adaptive design allows prospective decisions to be made based on accumulating results that become available during the course of the trial, such that more patients can be enrolled in cohorts showing the strongest efficacy signals and poorly performing cohorts can be closed early; and additional treatment cohorts testing novel agents and combinations may also be added. Although the clinical utility of genotype-matched trials remains to be proven, they provide a framework to efficiently probe the relationship between molecular aberrations, tumor histology, and MTA activity. Genotype-based trials may also be distinctly advantageous in matching targeted treatments for patients with low prevalence mutations, although extensive screening efforts or parallel molecular screening programs will be required to identify these patients. The establishment of national or international registries of patients whose tumors have undergone molecular profiling that enables rapid identification of rare genotypes might be useful. An additional challenge is that most of these molecularly based clinical trials rely on genomic sequencing data from a single tumor sample, offering only a static “snapshot” of patients’ genomic profiles. Cancer Research UK’s TRACERx trial sets a precedent for a prospective non-interventional study using multiregional and longitudinal tumor and circulating biomarker sampling to map the impact of dynamic intra-tumor heterogeneity during the course of disease progression. Insights gained from this study will inform the future development of dynamic therapeutic genomic trials [102].

To support molecular-based patient selection strategies, biomarker development and validation are crucial, preferably in concert with drug discovery and testing. A successful example is the development of crizotinib in anaplastic lymphoma kinase fusion gene (ALK)-translocated non-small cell lung cancer, in which the identification of a molecular subset of patients helped accelerate drug registration [4]. The currently ongoing I-SPY 2 (NCT01042379) is an adaptive multi-cohort trial for locally advanced breast cancer, aiming to identify biomarker-matched MTAs and MTA combinations. So far, three experimental therapies have demonstrated improved activity compared with standard therapy in distinct biomarker signature populations and have met pre-specified criteria for testing in confirmatory phase III trials [103–105]. In a recent review, biomarkers that aid patient selection were found to improve the phase transition probabilities of oncology drugs, although the majority of non-orphan drugs are still being developed without markers [106]. The US Institute of Medicine has released ten recommendations for the clinical development and use of biomarker tests, in recognition of the wide-ranging obstacles in the context of tumor heterogeneity, substantial technical difficulties in assay reproducibility and standardization, and reimbursement barriers [107]. Challenges notwithstanding, drug–biomarker pairing is vital to enhance the therapeutic index of MTAs and bring maximal benefit to the appropriate target population.

Precision medicine: the individualized dynamic model

In addition to spatial tumor heterogeneity, there is increasing awareness of tumor clonal evolution as a key mechanism of therapeutic failure, whereby genomic and epigenetic alterations and resistant variants develop and proliferate under selective treatment pressures [108, 109]. Thus, effective precision medicine would need not only to respond to the molecular diversity unique to each individual patient but also to adapt to the evolving dynamics of cancer. The individualized dynamic model may be a solution to this complex challenge, allowing intelligent drug combinations to be tailored to the genomic and immune profile of individual patients. Critical to this approach is the longitudinal monitoring of the changing molecular landscape to assess for treatment efficacy, to enable the early discovery of emerging resistant clones, and to target these pre-emptively with new drugs or combinations prior to the onset of clinical or radiological progression.

To facilitate the collection of dynamic molecular information, novel techniques such as serial measurements of circulating tumor DNA (ctDNA) or cell-free DNA (cfDNA) can be used at key treatment time points or regular intervals and present a less invasive alternative to tumor biopsies [110, 111]. A recent study of serial cfDNA sampling utilizing next-generation sequencing in phase I patients demonstrated the feasibility of this approach and suggested that cfDNA allele frequency dynamics may correlate with clonal response to targeted therapy [112]. Moreover, a number of other approaches are showing early promise to assist with dynamic therapeutic monitoring and prediction of specific treatment susceptibilities. For example, patient-derived xenografts using patients’ own avatars for drug sensitivity testing may help to predict the emergence of resistance clones ex vivo and inform therapeutic options, although clinical application requires successful engraftment and timely generation of models [113, 114]. The use of patient-derived organoids may provide a suitable alternative with a faster turnaround time. Emerging radiomic techniques, which enable high-throughput extraction of a large number of quantitative features from imaging modalities (computed tomography (CT), positron emission tomography (PET), or magnetic resonance imaging (MRI)), bring hope of providing noninvasive methods to track phenotypic changes in anatomical imaging during treatment, and associations with underlying gene expression patterns have been shown [115, 116].

The proposed individualized dynamic model offers the compelling potential to deliver immediate and durable benefit to patients, as well as an opportunity to study disease evolution and biology at an individual genetic level. However, the implementation and scaling-up of this approach may face a myriad of technical, resource, and culture issues. In Table 4, key considerations in designing and executing dynamic genomics trials are highlighted. Figure 2 presents an example of an individualized dynamic trial design.

Key components	Comments
Molecular and immune profiling at baseline	• Whole-exome or whole-genome sequencing, ideally using fresh tumor biopsies • Transcriptomic profiling • Proteomic validation • Immuno-phenotyping • Parallel ctDNA/cfDNA • PDX or PDO
Dynamic monitoring of molecular and immune landscapes	• Serial tumor biopsies may trigger concerns of safety and may not capture spatial heterogeneity • Novel strategies include serial monitoring of ctDNA/cfDNA, molecular imaging, and radiomics evaluation • ctDNA/cfDNA may assist with monitoring of treatment efficacy and early detection of resistant clones, although whether circulatory biomarkers reflect spatial tumor heterogeneity remains to be addressed • Assays need to have fast turnaround to allow “real-time” decision-making
Correlation of molecular monitoring with radiological response	• Requires exploration in future studies • Radiological response may not always provide the full picture, as in the case of mixed responses, and may lag behind treatment resistance • Also, pseudoprogression may occur in some cases with immuno-oncology agents
Multidimensional treatment algorithms at key decision points in response to molecular results	• If multiple mutations are present, treatment prioritization is required. Considerations may include relevance and level of evidence for the actionability of the mutation(s): that is, “driver” versus “passenger” mutations; allele frequency of mutation(s), and copy number change in the case of amplifications; downstream and parallel pathway aberrations that may confer treatment resistance; and availability of drugs and drug combinations. Sequential or alternating approaches may also be considered • Evaluation of immunotherapeutic approaches in cases of high mutational burden or other immune biomarkers to assess their predictive role • With the detection of emerging clones, consider changing therapeutic strategy ahead of radiological progression
Access to approved and investigational agents	• Requires collaboration with industry and academic partners

ctDNA circulating tumor DNA, cfDNA cell-free DNA, PDO patient-derived organoid, PDX patient-derived xenograft